Merge pull request #212 from PredicateSystems/pause_resume

agent pause/resume feature

Author: rcholic

src/agents/planner-executor/extraction-keywords.ts

@@ -41,6 +41,14 @@ const AMBIGUOUS_VERBS: readonly string[] = [
   'show',
   'tell',
   'display',
+  'provide',
+  'report',
+  'give',
+  'identify',
+  'collect',
+  'gather',
+  'return',
+  'output',
 ];
 
 /**
@@ -67,6 +75,14 @@ const EXTRACTION_PHRASES: readonly string[] = [
   'find the name',
   'how many',
   'how much',
+  'sale price',
+  'sale prices',
+  'first 5',
+  'first 10',
+  'first 3',
+  'top 5',
+  'top 10',
+  'top 3',
 ];
 
 /**
@@ -121,6 +137,9 @@ const CONTENT_NOUNS: readonly string[] = [
   'url',
   'image',
   'photo',
+  'product',
+  'results',
+  'listings',
 ];
 
 /**
@@ -267,29 +286,38 @@ export function getExtractionDomainGuidance(): string {
   return `
 
 IMPORTANT: Extraction Task Planning Rules
-=========================================
-For extraction tasks where data is already visible on the page:
+========================================
 
-1. If the data you need is VISIBLE in the page context above:
-   - Use EXTRACT directly as the ONLY step - no clicking needed
-   - The EXTRACT action will read the visible text from the page
+STEP 1 - CHECK CURRENT URL:
+Before choosing an action, compare the Current URL to the goal.
+- Does the current page contain the data requested?
+- If the goal mentions a specific section/page (e.g., "show hn", "top stories", "/show"), check if the URL matches.
+- If you are NOT on the right page, NAVIGATE to the correct URL first.
 
-2. If you need to navigate to see the data:
-   - First CLICK or NAVIGATE to the right page
-   - Then use EXTRACT
+STEP 2 - EXTRACT VISIBLE DATA:
+If the data is VISIBLE in the page context:
+- Use EXTRACT directly - no clicking needed
+- The EXTRACT action reads visible text from the current page
 
 CRITICAL: Do NOT click on links to external sites when extracting.
 - Post/article titles often link to EXTERNAL sites
 - To extract a title that is visible, use EXTRACT directly on the current page
 - Only click if you need to navigate to a detail page (e.g., for comments)
 
-Example for "Extract the title of the first post":
-{
-  "action": "EXTRACT",
-  "target": "first post title",
-  "goal": "Extract the first post title from the page",
-  "verify": []
-}
+Example - wrong page, need to navigate first:
+Goal: "extract the title of the first showhn post on hackernews show"
+Current URL: news.ycombinator.com/news (wrong page, need /show)
+{"action":"NAVIGATE","target":"https://news.ycombinator.com/show","verify":[{"predicate":"url_contains","args":["show"]}],"reasoning":"navigate to Show HN page"}
+
+Example - on correct page, extract directly:
+Goal: "extract the title of the first showhn post"
+Current URL: news.ycombinator.com/show (correct page, data visible)
+{"action":"EXTRACT","target":"first ShowHN post title","goal":"Extract the title of the first ShowHN post","verify":[],"reasoning":"data is visible on current page"}
+
+Example - product price on listing page:
+Goal: "find the price of the first laptop"
+Current URL: store.com/laptops (correct page, prices visible)
+{"action":"EXTRACT","target":"price of first laptop","goal":"Extract the price of the first laptop listing","verify":[],"reasoning":"prices are visible in listing elements"}
 `;
 }
 

src/agents/planner-executor/plan-models.ts

@@ -58,8 +58,8 @@ export type ActionType = z.infer<typeof ActionType>;
  * Type for a plan step.
  */
 export interface PlanStep {
-  id: number;
-  goal: string;
+  id?: number;
+  goal?: string;
   action: ActionType;
   target?: string;
   intent?: string;
@@ -87,12 +87,15 @@ const HeuristicHintSchema = z.object({
 
 export const PlanStepSchema = z.lazy(() =>
   z.object({
-    id: z.number().describe('Step ID (1-indexed, contiguous)'),
-    goal: z.string().describe('Human-readable goal for this step'),
+    id: z.number().optional().describe('Step ID (1-indexed, contiguous)'),
+    goal: z.string().optional().describe('Human-readable goal for this step'),
     action: ActionType.describe(
       'Action type: NAVIGATE, CLICK, TYPE, TYPE_AND_SUBMIT, SCROLL, PRESS, WAIT, EXTRACT, STUCK, DONE'
     ),
-    target: z.string().optional().describe('URL for NAVIGATE action'),
+    target: z
+      .union([z.string(), z.record(z.string(), z.unknown())])
+      .optional()
+      .describe('URL for NAVIGATE action'),
     intent: z.string().optional().describe('Intent hint for CLICK action'),
     input: z.string().optional().describe('Text for TYPE_AND_SUBMIT action'),
     verify: z.array(PredicateSpecSchema).default([]).describe('Verification predicates'),

src/agents/planner-executor/plan-utils.ts

@@ -477,6 +477,13 @@ function normalizeStep(step: Record<string, unknown>): Record<string, unknown> {
   if ('target' in normalizedStep && typeof normalizedStep.target === 'number') {
     normalizedStep.target = String(normalizedStep.target);
   }
+  if (
+    'target' in normalizedStep &&
+    typeof normalizedStep.target === 'object' &&
+    normalizedStep.target !== null
+  ) {
+    normalizedStep.target = JSON.stringify(normalizedStep.target);
+  }
   if ('target' in normalizedStep && normalizedStep.target === null) {
     delete normalizedStep.target;
   }
@@ -629,33 +636,36 @@ export function validatePlanSmoothness(plan: Plan): string[] {
 
   // Check each step
   let prevAction: string | null = null;
+  let prevId: number | undefined = undefined;
   for (const step of plan.steps) {
+    const stepLabel = step.id ?? '?';
     // Check for missing verification
     if ((!step.verify || step.verify.length === 0) && step.required !== false) {
-      warnings.push(`Step ${step.id} has no verification predicates`);
+      warnings.push(`Step ${stepLabel} has no verification predicates`);
     }
 
     // Check for consecutive same actions (might indicate loop)
     if (step.action === prevAction && step.action === 'CLICK') {
-      warnings.push(`Steps ${step.id - 1} and ${step.id} both use ${step.action}`);
+      warnings.push(`Steps ${prevId ?? '?'} and ${stepLabel} both use ${step.action}`);
     }
 
     // Check for NAVIGATE without target
     if (step.action === 'NAVIGATE' && !step.target) {
-      warnings.push(`Step ${step.id} is NAVIGATE but has no target URL`);
+      warnings.push(`Step ${stepLabel} is NAVIGATE but has no target URL`);
     }
 
     // Check for CLICK without intent
     if (step.action === 'CLICK' && !step.intent) {
-      warnings.push(`Step ${step.id} is CLICK but has no intent hint`);
+      warnings.push(`Step ${stepLabel} is CLICK but has no intent hint`);
     }
 
     // Check for TYPE_AND_SUBMIT without input
     if (step.action === 'TYPE_AND_SUBMIT' && !step.input) {
-      warnings.push(`Step ${step.id} is TYPE_AND_SUBMIT but has no input`);
+      warnings.push(`Step ${stepLabel} is TYPE_AND_SUBMIT but has no input`);
     }
 
     prevAction = step.action;
+    prevId = step.id;
   }
 
   return warnings;